}

func.func @cast(%arg0: tensor<1x2x2x5xi32>) -> tensor<1x2x2x5xf32> {
  %0 = "tf.Cast"(%arg0) : (tensor<1x2x2x5xi32>) -> tensor<1x2x2x5xf32>
  func.return %0 : tensor<1x2x2x5xf32>

  // CHECK-LABEL: cast
  // CHECK: "tfl.cast"(%arg0) : (tensor<1x2x2x5xi32>) -> tensor<1x2x2x5xf32>
}

func.func @castFloat32ToI16(%arg0: tensor<1x2x2x5xf32>) -> tensor<1x2x2x5xi16> {

  %0 = "mhlo.broadcast_in_dim"(%cst0) <{broadcast_dimensions = dense<[1, 3]> : tensor<2xi64>}> : (tensor<2x3xf32>) -> tensor<1x2x2x3xf32>
  %1 = mhlo.multiply %0, %cst1 : tensor<1x2x2x3xf32>
  // CHECK:      return %[[RES]] : tensor<1x2x2x3xf32>
  func.return %1 : tensor<1x2x2x3xf32>
}

// CHECK-LABEL: @foldBroadcastInDimBeforeMulOp_bcast_dim_1D_int

// CHECK-LABEL: testSpaceToDepth
func.func @testSpaceToDepthF32(%arg0: tensor<1x2x2x1xf32>) -> tensor<1x1x1x4xf32> {
  // CHECK: %[[ARG:.*]]: tensor<1x2x2x1xf32>
  // CHECK: "tfl.space_to_depth"(%[[ARG]]) <{block_size = 2 : i32}> : (tensor<1x2x2x1xf32>) -> tensor<1x1x1x4xf32>
  %0 = "tfl.space_to_depth"(%arg0) {block_size = 2: i32} : (tensor<1x2x2x1xf32>) -> tensor<1x1x1x4xf32>
  func.return %0 : tensor<1x1x1x4xf32>
}

// -----

    %13 = "tf.Mul"(%11, %12) : (tensor<1xf32>, tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32>
    %14 = "tf.Round"(%13) : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32>
    %15 = "tf.Cast"(%14) {Truncate = false} : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xi32>
    %16 = "tf.AddV2"(%15, %arg10) : (tensor<1x2x2x3xi32>, tensor<i32>) -> tensor<1x2x2x3xi32>

    : (tensor<?x3x2x1xi8>, tensor<2x1x1x1xi8>) -> tensor<?x2x2x1xi32>
  %1 = chlo.broadcast_add %0, %zp_offset : (
      tensor<?x2x2x1xi32>, tensor<?x2x2x1xi32>) -> tensor<?x2x2x1xi32>
  %2 = chlo.broadcast_add %1, %bias : (
      tensor<?x2x2x1xi32>, tensor<1xi32>) ->tensor<?x2x2x1xi32>
  %3 = chlo.broadcast_add %2, %bias : (
      tensor<?x2x2x1xi32>, tensor<1xi32>) ->tensor<?x2x2x1xi32>
  return %3 : tensor<?x2x2x1xi32>
}

// -----

// CHECK: return %[[DEQUANTIZED]] : tensor<*xf32>
}

// -----

module {
  func.func @not_quantize_matmul_without_const(%arg0: tensor<1x2x2x2xf32>, %arg1: tensor<2x1024xf32>) -> (tensor<*xf32>) {
    %arg0_identity = "tf.Identity"(%arg0) {device = ""} : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
    %arg1_identity = "tf.Identity"(%arg1) {device = ""} : (tensor<2x1024xf32>) -> tensor<2x1024xf32>

// function signature. Constants are also transposed accordingly.

// CHECK-LABEL: nchw_conv_with_bias_add
// CHECK-SAME: %[[ARG:.+]]: tensor<1x2x5x5xf32>
func.func @nchw_conv_with_bias_add(%arg0: tensor<1x2x5x5xf32>) -> tensor<1x4x5x5xf32> {
  %0 = stablehlo.constant dense<2.000000e+00> : tensor<4x2x3x3xf32>
  %1 = stablehlo.constant dense<3.000000e+00> : tensor<1x4x5x5xf32>

  %output = "tf.Cast"(%q_output) {Truncate = false} : (tensor<1x2x2x1x!tf_type.qint32>) -> tensor<1x2x2x1xi32>

  // CHECK-DAG: %[[OUTPUT_2:.*]] = "tf.Cast"(%[[OUTPUT_QINT]]) <{Truncate = false}> : (tensor<1x2x2x1x!tf_type.qint32>) -> tensor<1x2x2x1xi32>
  // CHECK-DAG: %[[OUTPUT_QINT_1:.*]] = "tf.Cast"(%[[OUTPUT_1]]) <{Truncate = false}> : (tensor<1x2x2x1xi32>) -> tensor<1x2x2x1x!tf_type.qint32>

      -> tensor<1x2x2x2xf32>
    %3 = "quantfork.stats"(%arg2) {layerStats = dense<[7.05456924, 7.11401462]> : tensor<2xf32>} : (tensor<2xf32>) -> tensor<2xf32>
    %4 = "quantfork.stats"(%2) {layerStats = dense<[-1.36523, 3.57373]> : tensor<2xf32>} : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
    %5 = "chlo.broadcast_add"(%4, %3) : (tensor<1x2x2x2xf32>, tensor<2xf32>) -> tensor<1x2x2x2xf32>

    } : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32>
    // CHECK-SKIP: tf.PartitionedCall
    // CHECK-NOSKIP: call @some_other_func
    // CHECK-NOSKIP-NOT: tf.PartitionedCall
    func.return %1: tensor<1x2x2x3xf32>
  }

  // CHECK-SKIP: func.func private @some_func
  func.func private @some_func(%arg0: tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32> attributes {tf._noinline = true} {